Take part

Lend your computer and the AI model of your choice. Your model reads a paper's title and abstract and writes, in two plain sentences, the problem it works on and how, without its field's words.

Fields often name the same problem differently: estimating a hidden state from noisy measurements is a Kalman filter to a control engineer and data assimilation to a weather forecaster. Plain descriptions let work on the same problem in different fields find each other. That takes one model call per paper, far more than one project can make, and easy when many people each make a few hundred.

Start

You need Python 3.10 or newer, a GitHub account with the GitHub command-line tool logged in once (gh auth login), and a model: one on your own computer, a free one in the cloud, or one you already pay for.

pip install "undiscovered @ git+https://github.com/nicolas-maman/undiscovered#subdirectory=clients/python"
undiscovered setup            # which model, and an optional OpenAlex key
undiscovered work --units 5   # five units of twenty papers

undiscovered work --dry-run does one unit and saves what it would send, without sending it. The client runs below normal priority and deletes its downloads when each unit is done. Full instructions: the client's README.

Which model

• On your computer, free and private: Ollama, LM Studio, llama.cpp, vLLM or Jan. undiscovered setup finds them by itself.
• Free in the cloud, no graphics card needed: Google Gemini, Groq, OpenRouter's free models or Mistral's free mode, each with a free key.
• One you already pay for: OpenAI, Anthropic, or anything that speaks the OpenAI API. The project pays for nothing and asks for nothing.

Nobody knows yet which models do this well. Every description records the model that wrote it, and each paper is described by two volunteers, so finding out is part of the work. How to connect each one, and what is sent where.

Progress

What the descriptions find, and how the models compare: Same problem, different field.

Who has helped

Models used so far

What happens to your work

• The client sends your descriptions as an issue under your GitHub account, so anyone can see who did what. Never a key.
• A job on the project checks each one within minutes: two short parts, and no more than two of the paper's own distinctive words, which the unit lists and your model is told to avoid. The issue is answered with what was accepted, or why not.
• Accepted work is stored openly, under CC0, on the project's data branch, with the name of the model that wrote it.
• The rules are in the protocol; any program that follows it, in any language, is a full participant.